Available online at www.sciencedirect.com
Procedia Computer Science 16 (2013) 870 – 877
Conference on Syst Eds.: C.J.J. Paredis, C. Bishop, D. Bodner, Georgia Institute of Technology, Atlanta, GA, March 19-22, 2013.
An improved user experience model with cumulative prospect theory Feng Zhou, Roger J. Jiao * The George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332-0405, USA
Abstract User experience (UX) design involves decision making for an optimal mix of product attributes to offer pleasurable UX. While the cognitive influences on human decision making have been well addressed, the affective elements for analyzing and simulating human perception on UX in the prevailing computational models are missing. In order to incorporate both affective and cognitive factors in the UX model for decision making, cumulative prospect theory (CPT) under two different affective states is studied. The least-square curve fitting technique is used to find multiple parameters involved in CPT. It successfully estimates parameters to represent different cognitive tendency and affective influence. A case study of the aircraft cabin interior design is illustrated to show the potential and feasibility of the proposed method.
© The Authors. Authors. Published Published by by Elsevier Elsevier B.V. B.V. © 2013 2013 The Selection and/or Selection and/or peer-review peer-review under under responsibility responsibility of of Georgia Georgia Institute Institute of ofTechnology. Technology Keywords: User experience modeling; affective-cognitive decision making; cumulative prospect theory
1. Introduction As so many products are no longer islands of their own to fulfill self-contained functionality, the most important success factor is probably user experience (UX), where multiple interdependent design attributes are considered as a consistent whole to create unique UX, and importantly, to achieve high economic value. This is deemed to be [1]. Human perception on UX originates (or emotional) states and cognitive processes) along the from evolution of chain of human-product interactions with choice decision making [2]. Engineering design traditionally copes with cognitive needs. Human emotional experience plays a significant role in decision making towards product success [3] [4]. While the cognitive influences on human decision making have been well addressed, the affective elements for analyzing and simulating human perception on UX in the prevailing computational models are missing. [5]. Expected utility theory assumes that humans make decisions on the basis of a deliberative cost-benefit analysis [6]. Recent models based on behavioral decision theories focus on cognitive errors and heuristics in human judgments and decision making, but still ignore the role of emotion in human decision making [7]. Such a single cognitive perspective is not optimal for analyz the time of decision making often influence their experience [8]. Recent consensus on the integration of emotion and
* Corresponding author. Tel.: +1-404-894-3256; fax: +1-404-894-9324. E-mail address:
[email protected].
1877-0509 © 2013 The Authors. Published by Elsevier B.V. Selection and/or peer-review under responsibility of Georgia Institute of Technology doi:10.1016/j.procs.2013.01.091
Feng Zhou and Roger J. Jiao / Procedia Computer Science 16 (2013) 870 – 877
871
cognition has been driven by the intimate coupling of affective and cognitive decisions [9]. Several computational mechanisms have emerged that treat cognition as a necessary antecedent to emotion [10]. However, the computational models of affect-cognition integration have largely been pragmatic [11], in that the link between cognitive functions and emotion has yet to be fully explored [12]. Traditional decision analysis seldom models explicitly the role of emotion in human decision making. Zajonc [13] argues that emotional reactions constitute the primary and determining response to social stimuli and consequently influence human judgments substantially. Damasio [14] demonstrates that without emotions, the cantly ability of decisionimpact the decision making process [15]. Therefore, it is necessary to develop affect-integrated decision models to describe human choice behavior and subjective experience under uncertainty. In addition, UX modeling and design necessitates identification of appropriate UX measures that coincide with customer preference and choice decision making, while incorporating the influence of cognitive tendency and affective states. 2. Related work 2.1 User-centered design We have been convinced by the trend of product value fulfillment progressing from traditional function-focused design to nowadays customization and personalization for UX design [16]. Pine and Gilmore [17] envision an This notion is consistent with user-centered design, which has been typically addressed in the fields of human-computer interaction, human factors and ergonomics, such as emotional design [18], Kansei engineering [19]. User-centered design concentrates more on the functionality and usability aspects of products and emphasizes affective perception of the product use, with little concern of how -centered design, such as naturalistic observation, protocol analysis, semantic differential, probes, narrative analysis, to name but a few, are largely qualitative and experiment based. This paper is geared towards a rigorous analytical approach by formulating UX design as a decision analysis problem that involves affective-cognitive decision making under uncertainty. 2.2 Preference modeling There exist various constructs, principles, and models formulated to predict and analyze customer preference and choice behavior in a variety of application contexts (e.g., [20]). In engineering design, quality function deployment (QFD) is one of the most commonly used methods to express customer preferences [21]. Usually a house of quality is formulated to map product features and functionality favored by customers. Some methods, such as fuzzy set [22], are incorporated to deal with the uncertainty involved. For instance, Mazur [12] makes use of QFD to translate lifestyle, image, and psychological needs into design requirements for designing the B787 Dreamliner commercial aircraft. Discrete choice experiments are also widely used to identify patterns in choices that customers make among competing products [23]. These methods can relate the choice made by the customer to the attributes of a product and other alternatives statistically. For example, in marketing, it allows for the examination of the interaction between market shares and product features, price, service, and promotion with respect to different classes of customers (e.g., [24]). Besides, conjoint analysis has proven to be an effective means to estimate part-worth utilities for individual product attributes [23]. These methods seek to identify optimal product concepts according to the -worth preference functions that are estimated within a conjoint framework (e.g., [25]). 2.3 Affect-cognition integration Storbeck and Clore [26] posit that affect and cognition are highly interdependent because the phenomena themselves are coupled. There is an increasing tendency to study the interaction between affect and cognition to understand the fundamental human needs for human subjective experience. For example, Lisetti and Nasoz [27] investigate how affect interacts with cognition and develop a multimodal affective user interface to simulate human intelligence. Ahn and Picard [8] propose an affective-cognitive decision framework for learning and decision making. Regardless of such common consensus, it is challenging to develop rigorous analytical tools to uncover
872
Feng Zhou and Roger J. Jiao / Procedia Computer Science 16 (2013) 870 – 877
customer affective and cognitive needs, to measure subjective experience, and to identify the mapping relationship between UX and design attributes. 3. Problem description (1) UX modeling: -cognitive decision making through their interactions with a variety of design attributes, denoted as a set, A {ai }I , where I is the total number of design attributes. These design attributes embody the key characteristics of a product or service system. Each design k Li where Li is the total attribute may assume a number of levels, either discrete or continuous, Ai* {aik* }I Li number of levels (instances) of ai , and k denotes the k-th level of ai . For example, the interior lighting color of an aircraft cabin can be a design attribute and may assume five attribute levels (e.g., blue, orange, green, pink, and particular configuration of design attribute levels, comprising a finite set, X {xik }I Li While xik indicates a quantitative measure of UX for a specific design attribute level, X is the aggregated measure of a holistic perception on UX for the entire design. With regard to various attribute levels as well as their associated costs, it is important that users are able to make wise choice decisions in light of their perceived UX. (2) Cognitive tendency and risk attitudes in choice decision making: CPT handles the probabilities attached to postulates that decision weights tend to overweigh small probabilities of design attributes and underweigh moderate and high probabilities of design attributes [28], which is referred to as probability distortion. In addition, a CPT value function is defined with respect to a reference point (stands for neutral UX), rather than absolute value as in expected utility theory. Such an emphasis on reference point conforms to the human perceptual process, which tends to notice shifts more than resting on static states [29]. The value function v exhibits a diminishing marginal value as the subjective value of UX moves further away from the reference point. This is referred to as diminishing sensitivity. In the pleasant UX domain, this implies concavity, i.e., v '' 0 ; and in the unpleasant UX domain, this implies convexity, i.e., v '' 0 (see Fig. 1). Since unpleasant UX looms larger than pleasant UX, it is characterized by the value function as steeper for unpleasant UX than for pleasant UX. (3) Affective influences on decision making: For social-psychological and economic decisions, Ahn [5] reveals that affective influence can be modeled through shape parameters of prospect theoretic value functions. Such an affective-cognitive model based on gains or losses of an economic outcome can hardly be directly applicable to economic decision making. We have to extend CPT to incorporate the unique characteristics of UX modeling. Also 4. Improved UX modeling The improved UX model consists of four phases, i.e., the perception, cognitive reasoning phase, shape fitting, and evaluation as described below: 4.1 Perception aik* with its value set Ai* can be defined as a subjective value function, v a . In the perception phase, the perceived UX of various options is identified relative to a certain design attribute level ai*, ref that gives a neutral UX and acts as a reference point. Hassenzahl and Tracinsky [30] point out that UX necessitates dynamic, context-dependent internal states of users, which involve both instrumental and emotional aspects. It is likely that the reference point varies among different respondents. To hedge against this problem, we set up individual reference points for individual UX models for customer heterogeneity and take a grand mean as the reference point for all the customers for customer homogeneity (in one customer segment). * ik
4.2 Cognitive reasoning
Feng Zhou and Roger J. Jiao / Procedia Computer Science 16 (2013) 870 – 877
873
(1) CPT subjective value function: According to CPT [29], the value function for a perceived attribute-level UX can be formulated as the following:
vik
aik*
* ik
v a
* ik
, aik*
,
0 * ik
,
(1)
0
aik* aik* ai*, ref , is the perceptual difference between the reference attribute level and the target design attribute level. In addition, and are free parameters that vary between 0 and 1, modulating the curvature of the subjective value function and indicating the risk attitude of the customer. For , the larger the value, the more riskseeking the customer tends to be. For , the larger the value, the more risk-averse the customer would be. Moreover, where
below the reference point, with larger values expressing more aversion and sensitivity to unpleasant UX. (2) Affective shaping: Fig. 1 shows the subjective value functions. The curve is maximally sensitive to change nearest to the reference and progressively less sensitive as it moves away from it. It shows that customers tend to be risk-averse in the pleasant UX domain and become risk-seeking in the unpleasant UX domain. Moreover, the function is steeper in the unpleasant UX domain than that in the pleasant UX domain, showing that people are more sensitive to unpleasant UX. The changing shape is modulated by , , and , reflecting the influence of affective states on cognitive tendency in choice decision making [5]. Perceptual ptual UX U
v
0
1
1
2 2
, more risk-seeking 1
, more rrisk-averse
v '' 0 (Designn attribute levels result inn unpleasaan annt UX)
v '' 0 0
1
2
(Designn aattribute levels Reference R f
resultlt in i pleasant ple l asa UX)
1
Steeper, >1 1
, more risk-seeking g
2
, more risk-averse
Fig. 1: CPT-based UX value functions
(3) Choice probability: Original formulation of CPT is motivated for economic outcomes and thus the choice behavior is crafted as a subjective probability by transforming the objective probability of an outcome using weight functions [29]. It is true that different economic outcomes occur with varying probabilities; but not for the choice behavior of UX design, whereby design attribute levels are always available for customers to choose. Therefore, modeling of CPT choice probabilities should be consistent with the customer choice behavior in product design. Quantitative modeling to predict choice is an established area of research in marketing [31] and product planning [32]. Using random utility discrete choice models, it is possible to predict customer preferences on different design attribute levels [33]. The utility of a design attribute level aik* to the customer is indicated by v(( ik* ) . We can construct a closed form of choice probability adapted from the popular logit model [34], i.e., pik
where
p aik*
exp
(
* ik
)
/
L k 1
exp
(
* ik
)
,
(2)
> 0 is a scaling parameter. As
t shares can be carried out subsequently to elaboration of preference estimation by post hoc optimization with respect to . (4) UX evaluation function: A design attribute ai with multiple levels, i.e., Ai* { ik* }I Li k Li, can be transformed into m n 1 UX outcomes of subjective value as perceived by one customer. Arrange the outcomes in vii00 viin , which occur with respective probabilities, p im ,..., pi 0 ,..., pin . Note an increasing order, i.e., v im
874
Feng Zhou and Roger J. Jiao / Procedia Computer Science 16 (2013) 870 – 877
that vi0 corresponds to the outcome of the reference level; those smaller than vi0 are related to the outcomes of unpleasant design attribute levels; and those larger than vi0 are attributed to the outcomes of pleasant attribute levels. The decision maker evaluates each UX outcome in conjunction with the associated choice probability, and thus an UX of ai can be defined as the following: vik pix , aik* 0 xik X aik* , vik pix , aik* 0
w
n
pix
w
x
pin
w
pix
j x
j
pij m
pij
pin ,
n
w
j x 1 x 1
w pi ,
m
pij , 0
j
w
m
x
n 1,
pij , 1 m
pi ,
m
x
(3)
0,
.
The weight function, w, should take the following form based on CPT [29]: w( pix ) pixz / ( pixz (1 pix ) z )1/ z ,
(4)
where 0 z 1 specifies the inverse s-shaped weight function, such that z stands for pleasant UX (i.e., w w ) and z suggests unpleasant UX (i.e., w w ). Decreasing the value of z makes the function become more curved and cross the 45-degree line further to the right. This function shows that customers tend to over-weigh low probabilities with extreme outcomes of attribute levels and underestimate moderate and high probabilities. One good example is that customers often overweigh the value of the first-class cabin with a low probability but underestimate the value of the economy cabin with a high probability. 4.3 Shape fitting To support UX evaluation, it is necessary to find the shape parameters using curve fitting techniques under different affective states. In this paper, the least-square fitting technique is used to find the parameters involved in the CPT choice model. Since we have the CPT-based value function and weight function, the parameters are found by fitting the selfreport data to these functions. The lest-square fitting technique [35] is a simple and most commonly used one to find the best fitting curve to a given set of points by minimizing the sum of the squares of the residuals of the points from between 0 and 1, and and the curve. There are six parameters involved in the CPT-based UX model, i.e., , 1)-(4)). The exponential functions of Equations (1), (2) and (4) are linearized and linear least squares fitting is applied iteratively to a linearized form of the function until convergence is achieved. 4.4 Evaluation Once we obtain the parameters of CPT, it is possible to measure profile with specific design attributes ai 1 i I by X (a, p)
I i 1
v aik*
pix
of a design I
x .
i 1 ik
(5)
Note xik is a quantitative measure of UX for a specific design attribute level, whereas X is the aggregated measure weighted sum of individual xik in Equation (3). According to Equation (5), under CPT, a UX prospect X (a1 , p) , is preferred to another prospect, X (a2 , p) for a specific customer if X (a1 , p) X (a2 , p) and is indifferent if X (a1 , p) X (a2 , p) . Based on the descriptions above, preferences in terms of UX are determined jointly by a subjective value function that evaluates individual UX of specific design attribute levels with regard to a reference point, and by the decision weights that capture an nd affective influences, which make the weighted sum of individual xik a reasonable mental process. Under the circumstances of product design, for multiple design attributes A {ai }I , each with several levels Ai* {aik* }I Li , the improved UX model with CPT can be used to evaluate alternative design profiles in the design space.
875
Feng Zhou and Roger J. Jiao / Procedia Computer Science 16 (2013) 870 – 877
5. Case Study This case study focuses on the aircraft cabin design, and it aims to create positive UX in the aircraft cabin, including a healthier and more comfortable cabin environment. For illustrative simplicity without losing essence, the design attributes and their attribute levels are shown in Table 1 for the purpose of experimentation. (1) Participants: Twenty participants (10 Chinese and 10 Americans) with gender balance are recruited from Georgia Institute of Technology for the experiment. In order to increase homogeneity within cultural groups, Chinese students are required to be born and raised in mainland China, and to have lived in USA for less than 3 years. All the participants are aged between 20 and 30. Other related data, such as travel frequency and geographic information, are also collected. Informed consent is obtained from each participant. (2) Cabin Environment: The cabin environment is built in a virtual environment using VisionStation display. Compared to typical displays, it takes peripheral vision into account and thus it has a very wide field of view. The experience when watching such a display is very natural and smooth with a sense of real 3D motion and no distractions. Therefore, the participants can be immersed into the 3D virtual environment where the behavior of users is better contextualized. It is generally helpful to achieve reliable and efficient user navigation through the whole cabin environment and consequently contributes to an ecologically validated product-service system. (3) Data Collection: Before data collection, half of the participants are first required to watch a video clip to elicit fear and show how unpleasant they are. The second half participants are required to watch a video clip to elicit amusement and show how pleasant they are. The video clips are the boy playing in the hallway in the movie The Shining (eliciting fear) and the restaurant scene from the movie Drop Dead Fred (eliciting amusement) (see [36]). They are proved to successfully elicit the target affective states. The UX outcome is measured on a scale between 100 (extremely unpleasant) and 100 (extremely pleasant) for individual design attribute levels using a GUI developed with Matlab. Furthermore, they are required to make decisions between two presented design profiles. Therefore, each participant is required to make 26 decisions of all the design profiles and altogether 520 decision data are collected for 20 participants. Table 1: Aircraft Cabin design profiles for evaluations
Profile
Interior Color
Personal Space
Noise
Lighting
Interior Pattern
Basic Adjustable
Sculpted Ceiling with Gentle Curves
Not Adjustable Not Adjustable
Hard Lines or Flat Surface Sculpted Ceiling with Gentle Curves
1
Green
Restricted
Medium
26
White
Restricted
Medium
27
Green
Adequate
Medium
Air pressure Humidity Vibration Contaminant
Cost ($)
Low
20%-30%
Weak
Low
1200
6. Results The main goal of the study is to confirm affective shaping and cognitive tendency on the CPT-based UX parameters, and two data sets are used for two different affective states, i.e., fear as a typical negative state and amusement as a typical positive state. The decision data are randomly divided into 200 entries for the amused and the fearful participants, respectively, for parameter estimation and the remaining 60 entries for each type of participants are used for UX prediction. This process runs three times to generate the (averaged) results shown below. Table 2 shows the mean values and standard deviations of the estimated parameters for two different affective and for amusement are significantly larger than those for fear states. First, it is found that the mean values of ( : t(18) = 2.25, p < 0.05; : t(18) = 210.42, p < 0.001). It demonstrates that customers in the fearful condition tend to be more risk-averse than those in the amused condition both in the pleasant UX and unpleasant UX domains. These results are consistent with previous studies [37] that fear often coupled uncertain situations and low control tends to provoke more pessimistic risks whereas positive affective states, such as amusement, associate customers with optimistic expectations and thus they tend to be risk-seeking when the risk is low and tend to be risk-averse in for amusement and order to sustain the positive affective states when the risk is high. Second, the mean values of
876
Feng Zhou and Roger J. Jiao / Procedia Computer Science 16 (2013) 870 – 877
that for fear are not significantly different ( : t(18) = -1.86, p < 0.10). This is probably because unpleasant UX is so strong involved in the aviation industry that both of those types of customers are equally sensitive and aversion to it. Both for amusement and fear, participants over-weigh low probabilities with extreme outcomes and under-weigh high probabilities with extreme outcomes and are relatively insensitive to probability difference in the middle. Third, the mean values of ( : t(18) = -138.58, p < 0.001) and of ( : t(18) = -11.02, p < 0.001) are significantly smaller (i.e., more curved) for amusement than their counterparts for fear. This means that fearful individuals are more sensitive to extreme outcomes and less sensitive to moderate outcomes for both pleasant UX and unpleasant for two types UX than amused customers. Besides the parameters involved in CTP, the estimated mean values of of customers are also listed in Table 2. Table 2: Results of parameter estimation in two different affective states Parameter mean (standard deviation)
Affective states Amusement Fear
0.43 (0.17) 0.36 (0.15)
0.77 (0.02) 0.55 (0.03)
2.63 (0.95) 2.91 (0.98)
0.30 (0.03) 0.37 (0.02)
0.19 (0.10) 0.27 (0.14)
1.42 (0.68) 2.36 (0.82)
Based on these results, we can obtain the UX prediction function for any design profiles from the design space. For example, for the design profiles 1 and 3 in Table 1, their respective reference points obtained from the grand ons 1 mean are shown in Table 3 and 3 with regard to design profiles 1 and 3), we can quantify the UX using Equations (1) to (5) with the parameters estimated. The quan design options, design profile 1 will be chosen. Furthermore, the contribution of individual design attributes that lead to pleasant and/or unpleasant UX can also be specified. For example, in Profile 1, the personal space is restricted, accounting for 68.7% of unpleasant UX while the cost < 800 USD, accounting for 50.5% of pleasant UX. Table 3: UX comparison between design profile 1 and profile 3 -
Interior Color
Personal Space
Noise
Profile 1
Green
Restricted
Medium
Profile 3
Orange
Adequate
Low
Reference Evaluation1 Evaluation3
2.93 9.65 -1.70
-7.34 -35.94 0.65
-5.66 -4.38 12.13
Lighting
Interior Pattern
Basic Adjustable Basic Adjustable 1.17 3.69 3.69
Sculpted Ceiling with Gentle Curves Sculpted Ceiling with Gentle Curves -0.54 11.13 11.13
Air pressure
Humidity Vibration Contaminant
Cost ($)
Low
20%-30%
Weak
Low
